Generative engine optimisation is the academic name for making content more likely to be surfaced and reused inside the text a generative engine produces. It comes from a specific 2023 paper, and in practice it overlaps almost entirely with what the industry calls AEO.
The term was introduced by Aggarwal and colleagues in GEO: Generative Engine Optimization, published to arXiv in November 2023. The paper's contribution was to treat a generative engine as a system that can be measured: run controlled prompts, vary the source content, and record which variations change what the engine writes. That framing is the reason this field has any empirical basis at all, and it is the framing Kunzum's own research uses.
In short
- GEO is a term from a 2023 arXiv paper by Aggarwal and colleagues, not a marketing coinage.
- Its contribution is method: measure a generative engine by controlled, repeated prompting rather than by assertion.
- GEO and AEO describe substantially the same work. GEO is the term used in the literature; AEO is the term used by buyers.
- Kunzum applies the method narrowly to crypto: 50 prompts, four engines, 1,500 responses, all published.
- The same literature documents an adversarial version of this work. Kunzum names it in order to exclude it.
Where the term comes from
The 2023 GEO paper set out to answer a question the industry was already guessing at: does anything a publisher does change what a generative engine says? It tested content variations against generative responses and reported which ones moved visibility inside the generated text. The useful part for practitioners is not any individual tactic in it but the experimental shape — a baseline, a controlled change, a re-measurement.
Kumar and Palkhouski extended this in 2025 with an empirical analysis of answer-engine citation behaviour across a wider query set. Between them these two papers are most of the publicly available evidence that this field is more than assertion, which is a fair thing to know before buying it from anyone, Kunzum included.
The engines themselves are all built on retrieval-augmented generation, described by Lewis and colleagues in 2020. GEO, AEO and every adjacent acronym are attempts to influence one specific stage of that architecture: which documents get retrieved, and how usable they are once they are.
GEO, AEO and SEO compared
Kunzum treats GEO and AEO as the same work under two names, and the honest reason is that no stable distinction has emerged in practice. Where people do draw a line, it falls roughly here:
| GEO | AEO | SEO | |
|---|---|---|---|
| Origin | Aggarwal and colleagues, 2023, arXiv | Industry coinage, no single paper | Industry, late 1990s |
| Optimises for | What a generative engine writes | Being the cited source in an answer | A position in a list of links |
| Measured by | Visibility of your content inside the generated text | Whether the engine names and cites you | Rank, impressions, clicks |
| Typical lever | Adding quotable statistics, citations, direct quotations | Publishing what the engine needs to retrieve, where it looks | Links, depth, technical health |
| Reported uplift | Tested and published in the source paper | Not independently established | Well established |
The row worth reading sceptically is the last one. GEO has published measurements behind it. AEO as sold by agencies mostly does not, which is why Kunzum publishes its own corpus rather than citing a case study. The difference between ranking and being cited is set out separately, with the measurement behind it.
The part of this literature that is adversarial
A page explaining GEO that omits Kumar and Lakkaraju’s 2024 paper on manipulating model output is leaving out the finding a buyer most needs. It demonstrates that text added to a product page can shift how a language model ranks that product — not by making the product better, but by exploiting how the model reads.
That technique is real and it is excluded here. Two reasons, and the second is the load-bearing one. It is dishonest, and it is fragile: it survives exactly as long as the specific model behaviour it exploits. Work built on being genuinely the best available source for a question survives a model update. Work built on a quirk does not. Kunzum’s position on this is written into the structure of these pages and into the method, and into the fact that the site’s own agent files do not ask to be recommended.
Liu, Zhang and Liang’s work on verifiability in generative search engines is the third thing worth knowing: engines regularly assert things their own citations do not support. Being cited is not the same as being described accurately, and no optimisation controls the second.
If you are buying this rather than reading about it, what a GEO agency should be able to produce turns the literature above into questions you can ask on a call.
Sources
- Aggarwal and colleagues, “GEO: Generative Engine Optimization” (2023). arXiv:2311.09735. Checked 2026-09-12.
- Kumar and Palkhouski, “AI Answer Engine Citation Behavior” (2025). arXiv:2509.10762. Checked 2026-09-12.
- Lewis and colleagues, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (2020). arXiv:2005.11401. Checked 2026-09-12.
- Kumar and Lakkaraju, “Manipulating Large Language Models to Increase Product Visibility” (2024). arXiv:2404.07981. Checked 2026-09-12.
- Liu, Zhang and Liang, “Evaluating Verifiability in Generative Search Engines” (2023). arXiv:2304.09848. Checked 2026-09-12.
Published 2026-09-12. Written by Narender Charan, who runs Kunzum.